GitLab

* add v2 flux examples

* add the docs

* add docs

* update

* finished ops

* add ops

* update

* update

* update

* update

* update

* update

* update

* update docstrings

* update

* update

* update

* update

* update

* update

* update

* finished the api docs

* update

* update

* tmp

* update

* update

* finished the offloading impl

* the offloading is buggy

* update utils

* the offloading is still buggy

* update

* correctness and speedup done; need to check the vram overhead

* done

* final debugging

* update

* update

* correct now

* fix

* update

* use per-layer offloading

* fix the offloading on 5090

* support setting the num_blocks_on_gpu

* change the import name

* feat: add support for python 3.13

* update workflow to exclude py3.13+torch2.5

* fix typo

Co-authored-by: Kung Talon <31659820+kungtalon@users.noreply.github.com>

* Add paddings if LoRA ranks for q, k, v are different

* No need to create a list

* feat: add support for teacache in flux kontext

* merge main and make linter happy

* caching_v2

* rename fb cache and write docstring

* lint

* rename utils to fbcache

* no need maintain sana for caching

* update

* update

* update README

* update dos

* update docs

* improve the lightning script

* update the example script

* change the repo name

* update

* add parameter of act unsigned

* upgrade the diffusers to v0.35.1

* bump the comfyui version to 0.3.51

* update version

* revert the test comfyui-version back to 0.3.44

* fix: enhance latent variable preparation in NunchakuQwenImagePipeline

- Refactored latent variable preparation to utilize the parent method for generating latents.
- Added manual generation of latent_image_ids to ensure correct indexing.

* refactor: clean up whitespace in NunchakuQwenImagePipeline

- Removed unnecessary whitespace in the latent variable preparation section for improved readability.

* update app

* depth supports fp4

* update

* fix the demo website

* style: make linter happy

* start refract the codebase

* update

* update

* start to implement ops

* add gemm

* write the docstrings

* define the w4a4 svdq linear

* update

* make the linter happy

* finished the SVDQW4A4Linear

* finished the SVDQW4A4Linear

* update

* update

* add a patcher to the model

* update

* add adanormsinglezero

* update

* update

* finished the naive implementation of nunchaku flux

* add ff

* finished the naive forward

* update

* svdq linear

* start debugging

* fix some issues

* successfully built the model

* update

* successfully load the model

* update

* update

* update

* try to making it runnable

* debugging

* debugging

* debugging

* add bias to awq linear

* run through

* fix the normalization

* update

* update

* update

* fix the attention

* fix the no fuse nvfp models

* update

* finished the fused ff

* make linter happy

* make linter happy

* make linter happy

* debugging the fp16 attn

* nunchaku fp16 is buggy

* finish the fp16 attn

* fp4 done

* fix the lora scales

* add a default value for alpha; need to debug int4

* fix input4

* update

* update

* ff does not work

* specialize the processors

* qwen transformer done. start debugging

* make linter happy

* add schnell v2 for metrics eval

* chore: schnellv2 eval

* update

* ff and attention correct

* need to check what happened to module

* fp4 done

* make linter happy

* update an example script

* reformat

* add an example script

* add the annoucement

* remove a misleading info

* ready to release

* fix teacache_batch

* lint

* fix the hf uploading

* use conda python

* chore: support uploading wheel to hf

* support ready_for_review

* Fix FLUX.1-Kontext LoRA support and dimension mismatch issues

- Added convert_keys_to_diffusers() for ComfyUI/PEFT format conversion
- Fixed dimension mismatch in LoRA weight concatenation
- Added preprocessing for single_blocks LoRA structure
- Added comprehensive test suite for Kontext LoRA
- Added example script for FLUX.1-Kontext with LoRA

Fixes #354

* lint

* FAL.AI and relight-kontext-lora patch

* fix workflow for torch 2.8

* update based on main

* chore: add test trials

* update the test score

* fix: Re-enable transformer caching in apply_cache_on_pipe

* to pass tests

* add the doc string back

* add the doc string back

---------
Co-authored-by: Muyang Li <lmxyy1999@foxmail.com>